tensorflow 정확도가 NAN으로 바뀌 었습니다. ADAMOPT이 적용된 Google 자습서에서 적용

Question

mnist cnn tutorial을 SVNN 데이터에 적용하려고했습니다. 나는 오류를 얻지 못한다. (무언가가 무시당하는 것 외에) 그러나 훈련 정확도는 몇 배치 후에 NAN으로 간다. 저는 cnn과 tensorflow에 익숙하지 않아서, 왜 그렇게 길을 잃었습니다.

import urllib 
import os.path 
import numpy as np 
testfile = urllib.URLopener() 
testfile2=urllib.URLopener() 
import scipy.io as scp 

if not os.path.isfile("test.mat"): 
    testfile.retrieve("http://ufldl.stanford.edu/housenumbers/test_32x32.mat", "test.mat") 

if not os.path.isfile("train.mat"): 
    testfile.retrieve("http://ufldl.stanford.edu/housenumbers/train_32x32.mat", "train.mat") 
testdata=scp.loadmat('test.mat') 
traindata=scp.loadmat('train.mat') 
trainDataX = traindata['X'] 
trainDataY = traindata['y'] 
testDataX = testdata['X'] 
testDataY = testdata['y'] 
def OnehotEndoding(Y): 
    Ytr=[] 
    for el in Y: 
     temp=np.zeros(10) 
     if el==10: 
      temp[0]=1 
     elif el==1: 
      temp[1]=1 
     elif el==2: 
      temp[2]=1 
     elif el==3: 
      temp[3]=1 
     elif el==4: 
      temp[4]=1 
     elif temp[5]==1: 
      temp[5]=1 
     elif temp[6]==1: 
      temp[6]=1 
     elif temp[7]==1: 
      temp[7]=1 
     elif temp[8]==1: 
      temp[8]=1 
     elif temp[9]==1: 
      temp[9]=1 
     Ytr.append(temp) 

    return np.asarray(Ytr) 
trainDataY = OnehotEndoding(trainDataY) 
testDataY = OnehotEndoding(testDataY) 
def transposeArray(data): 
    print 'started' 
    xtrain = [] 
    trainLen = data.shape[3] 
    print trainLen 
    for x in xrange(trainLen): 
     xtrain.append(data[:,:,:,x]) 

    xtrain = np.asarray(xtrain) 
    return xtrain 
trainDataX = transposeArray(trainDataX) 
testDataX = transposeArray(testDataX) 
print trainDataX.shape 
import tensorflow as tf 
sess = tf.InteractiveSession() 


x = tf.placeholder(tf.float32, shape=[None, 32,32,3]) 

y_ = tf.placeholder(tf.float32, shape=[None, 10]) 

def weight_variable(shape): 
    initial = tf.truncated_normal(shape, stddev=0.1) 
    return tf.Variable(initial) 

def bias_variable(shape): 
    initial = tf.constant(0.1, shape=shape) 
    return tf.Variable(initial) 

def conv2d(x, W): 
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') 

def max_pool_2x2(x): 
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], 
         strides=[1, 2, 2, 1], padding='SAME') 

W_conv1 = weight_variable([5, 5, 3, 32]) 
b_conv1 = bias_variable([32]) 
x_image = tf.reshape(x, [-1,32,32,3]) 
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) 
h_pool1 = max_pool_2x2(h_conv1) 
W_conv2 = weight_variable([5, 5, 32, 64]) 
b_conv2 = bias_variable([64]) 
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) 
h_pool2 = max_pool_2x2(h_conv2) 
W_fc1 = weight_variable([8 * 8* 64, 1024]) 
b_fc1 = bias_variable([1024]) 
h_pool2_flat = tf.reshape(h_pool2, [-1, 8*8*64]) 
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) 
keep_prob = tf.placeholder(tf.float32) 
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) 
W_fc2 = weight_variable([1024, 10]) 
b_fc2 = bias_variable([10]) 
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) 


cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1])) 


train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) 
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1)) 
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) 
with tf.Session() as sess: 
    epoch=10000 
    batch_size=100 
    sess.run(tf.initialize_all_variables()) 
    p = np.random.permutation(range(len(trainDataX))) 
    trX, trY = trainDataX[p], trainDataY[p] 
    print len(trainDataX) 
    start = 0 
    end = 0 
    for step in range(epoch): 
     start = end 
     end = start + batch_size 

     if start >= len(trainDataX): 
      start = 0 
      end = start + batch_size 

     if end >= len(trainDataX): 
      end = len(trainDataX) - 1 
     inX, outY = trX[start:end], trY[start:end] 
     #sess.run(optimizer, feed_dict= {tf_X: inX, tf_Y: outY, keep_prob:0.75}) 
     if step % 100 == 0: 
      train_accuracy = accuracy.eval(feed_dict={x: inX, y_: outY, keep_prob:1}) 
      print("step %d, training accuracy %g"%(step, train_accuracy)) 
     train_step.run(feed_dict={x: inX, y_: outY, keep_prob: 0.5}) 

    print("test accuracy %g"%accuracy.eval(feed_dict={ 
    x: testDataX, y_:testDataY , keep_prob: 1.0})) 

Answer 1

크로스 엔트로피 수식이 잘못된 것 같다 : (i는 아담 최적화 useed과 문서가 동적으로 학습 속도를 적응 말했다 이후는, 학습 속도가 될 수 없습니다 IMHO) 는

여기 내 코드입니다. 대신이 tf.nn.softmax_cross_entropy_with_logits 내장 사용

logits = tf.matmul(h_fc1_drop, W_fc2) + b_fc2 
y_conv = tf.nn.softmax(logits) 

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, y_))